Scaling the profile of life by function with SPIN

Motivations: Classifying hundreds of thousands of protein sequences by function remains a significant computational challenge. Building on the ProfileView method for identifying functional classes and subclasses, our goal is to achieve large-scale classification of proteins from extensive databases and ongoing high-throughput sequencing efforts, ultimately producing comprehensive sets of sequences that share the same function. Results: By applying deep learning techniques, SPIN learns discriminative patterns in functionally related sequences, allowing the classification of hundreds of thousands of sequences into a defined number of functional classes. SPIN offers an effective compromise between small, family-specific protein language models (pLMs) and computational cost, with a time complexity linear in the number of sequences. It enables the identification of family-specific conserved residues, providing insight into the functional nuances of protein subclasses. By enhancing the scalability of protein function predictors, SPIN advances our understanding of protein functions and their evolutionary relationships. Availability and Implementation: The data and code that support the findings of this study are publicly available at https://gitlab.lcqb.upmc.fr/andrea.mancini/SPIN.